LLMs & Foundation Models

Anthropic Releases Claude Fable 5, Its First Public Mythos-Class Model

Fable 5 launched June 9 with guardrails that hand off cybersecurity and biology queries to Opus 4.8.

Liza Chan
Liza ChanAI & Emerging Tech Correspondent
June 9, 20263 min read
Share:
Abstract visualization of a powerful AI model with a protective barrier separating it from restricted zones

Anthropic released Claude Fable 5 on Tuesday, June 9, 2026, calling it the first publicly available model in its Mythos class. It arrives two months after the company unveiled Mythos itself, the model it was too nervous to ship.

What changed since April

Mythos got a preview in April and then stayed locked up. Anthropic handed it to a few partners through Project Glasswing, citing the model's knack for finding and chaining software exploits. So the obvious question was always how a company that spent two months warning people about a model would turn around and sell it.

The answer is safeguards and a fallback. Fable 5 is built on the same training as Mythos but ships with blocks on high-risk topics. Ask it something in cybersecurity, biology, chemistry, or distillation and you get routed to Claude Opus 4.8 instead, which was the strongest public model right up until Tuesday morning.

Whether that holds is the interesting part. Anthropic says its red team ran more than 1,000 hours against the guardrails and found no universal jailbreak. No universal jailbreak is not the same as no jailbreak, and the company seems to know it, which is presumably why the system card exists.

The numbers, and what they leave out

Anthropic claims Fable 5 is state-of-the-art on nearly all tested benchmarks, with the lead widening on longer tasks. One reported figure stands out: 95 percent of Fable sessions in testing ran entirely on Fable without falling back to Opus. That cuts two ways. It means the guardrails rarely fire, which is good for the 95 percent. It also means roughly one in twenty sessions quietly drops to a weaker model, and the company is not saying which kinds of work land in that bucket.

"Fable 5's capabilities exceed those of any model we've ever made generally available," the company wrote, which is the sort of line every lab writes about every release. The benchmark scores at least back the boast on coding.

Pricing landed lower than the rumors suggested. Early reports floated double the Opus tier; the actual numbers came in at $10 per million input tokens and $50 per million output, less than half what Mythos Preview cost. Subscribers get free access until June 22, after which the meter starts.

Mythos 5 stays behind the curtain

Alongside Fable, Anthropic shipped Mythos 5, basically the same model with the safety routing stripped out. That one is not for sale. It goes to the cyberdefenders and infrastructure providers who already had Mythos Preview access, including the Project Glasswing partners.

The split is the whole strategy. Sell the capable model to everyone, keep the dangerous version where it can be watched. The timing is hard to ignore: the launch follows Anthropic's own public warning that frontier systems are moving fast enough to approach recursive self-improvement, and it comes as the company eyes public markets.

Fable 5 is available now through the Anthropic API and consumption-based Enterprise plans. Free subscriber access runs through June 22.

Tags:AnthropicClaudeFable 5MythosAI modelsAI safetycybersecuritylarge language modelsClaude Opus
Liza Chan

Liza Chan

AI & Emerging Tech Correspondent

Liza covers the rapidly evolving world of artificial intelligence, from breakthroughs in research labs to real-world applications reshaping industries. With a background in computer science and journalism, she translates complex technical developments into accessible insights for curious readers.

Related Articles

Stay Ahead of the AI Curve

Get the latest AI news, reviews, and deals delivered straight to your inbox. Join 100,000+ AI enthusiasts.

By subscribing, you agree to our Privacy Policy. Unsubscribe anytime.